Intelligent Care Management for Diabetic Foot Ulcers: A Scoping Review of Computer Vision and Machine Learning Techniques and Applications

Abstract

Ten percent of adults in the United States have a diagnosis of diabetes and up to a third of these individuals will develop a diabetic foot ulcer (DFU) in their lifetime. Of those who develop a DFU, a fifth will ultimately require amputation with a mortality rate of up to 70% within five years. The human suffering, economic burden, and disproportionate impact of diabetes on communities of color has led to increasing interest in the use of computer vision (CV) and machine learning (ML) techniques to aid the detection, characterization, monitoring, and even prediction of DFUs. Remote monitoring and automated classification are expected to revolutionize wound care by allowing patients to self-monitor their wound pathology, assist in the remote triaging of patients by clinicians, and allow for more immediate interventions when necessary. This scoping review provides an overview of applicable CV and ML techniques. This includes automated CV methods developed for remote assessment of wound photographs, as well as predictive ML algorithms that leverage heterogeneous data streams. We discuss the benefits of such applications and the role they may play in diabetic foot care moving forward. We highlight both the need for, and possibilities of, computational sensing systems to improve diabetic foot care and bring greater knowledge to patients in need.

Keywords

computer vision diabetes complications diabetes management diabetic foot machine learning predictive models

Introduction

Ten percent of adults in the United States have a diagnosis of diabetes,¹ the leading cause of health care’s rising costs in the United States.² Over a third of those with diabetes will develop a diabetic foot ulcer (DFU) over their lifetime, frequently leading to infections and amputations with a mortality rate of up to 70% within 5 years.³ Communities of color are disproportionately impacted as they are more likely to develop this devastating complication of diabetes.^4,5

There is increasing interest in the use of computer vision (CV) and machine learning (ML) techniques to aid the detection, characterization, monitoring, and even prediction of DFUs. Diabetes-related limb loss pathophysiology is complex and highly variable, resulting in several challenges to successful diabetic foot care.³ Computer vision and ML may aid decision-making processes during the point of care as well as improve the continuity of care and patient engagement. This scoping review provides an overview of the benefits of interventions that incorporate CV and ML and the role they may play in diabetic foot care moving forward. We then provide an overview of relevant techniques for analysis of the at-risk foot, gait analysis, and predictive algorithms that leverage such heterogeneous data streams. This scoping review intends to bring attention to both the need for, and possibilities of, computational sensing systems which leverage CV and predictive ML algorithms to improve diabetic foot care.

Background

Diabetes and DFUs

Approximately 19% to 34% of patients with diabetes develop a DFU in their lifetime.³ Of those who develop a DFU, 40% will have recurrent ulcers within a year after ulcer healing, 50% will develop infections, and 20% will ultimately require amputation with a mortality rate of up to 70% within five years.³ Communities of color in the United States experience more frequent and severe diabetes complications including DFU.^4-6 In 2019, for example, non-Hispanic Black individuals were 2.5 times more likely to be hospitalized with diabetes complications, as compared to non-Hispanic white individuals.⁴ The cost in human suffering also translates to economic burden, as diabetes is the leading cause of the rising cost of health care in the United States.² Foot complications constitute a major source of the nation’s estimated annual diabetes expenditure, which totals $273 billion in direct costs and $90 billion in indirect costs.⁷

Diabetes-related limb loss pathophysiology is complex with variable clinical presentation. Due to frequent foot ulcer recurrence, management and treatment benefit from early and continuous assessment by foot experts such as podiatrists. Successful management of DFUs requires close coordination between health care team and patient. Like infection, critical limb ischemia is an important indicator for both DFU occurrence and recurrence, with high risks of amputation or death.⁸ Some have proposed the Comprehensive Foot Exam, also known as the Full Foot Exam, as a standard assessment for DFU management.⁹ The Full Foot Exam provides a quick assessment of a patient’s risk of foot ulceration by evaluating four key domains: dermatologic, musculoskeletal, neurological, and vascular.^9-11 In current practice, patients with DFUs are typically seen in the clinic in one- to four-week intervals to assess and treat their wound(s). It is therefore essential that the patients themselves are able to monitor their foot health in between visits, including changes in severity or recurrence of foot ulcers and presence of preulcerative lesions.

CV and ML

Machine learning is a subfield of artificial intelligence which aims to develop methods that identify patterns within data to “learn,” that is, to iteratively improve performance for some task.¹² Computer vision applications of ML models may train computers to identify, interpret, and/or extract information from digital images often in ways inspired by a human’s biological vision system. Applications of CV and ML suggest a range of promising solutions that could be leveraged to benefit the field of diabetic foot care, as well as wound care more generally.

The costs associated with DFUs, both economic and that of human suffering, highlight the need for scalable diabetic limb-preservation programs⁴ which emphasize multidisciplinary approaches focused on screening, remission, and engaging patients in self-monitoring. While there are disparate interventions and digital health technologies to track DFUs,^13,14 diabetic foot care may benefit from the development of a ubiquitous computational approach to foot health which includes monitoring before, during, and after wounds and other threats to the limb occur. There are many existing techniques in CV and ML that could be leveraged in support of this goal, including techniques for automated wound segmentation^15,16 and predictions of DFU progression.^17,18

Care in Place

Studies of the interplay between the processes of place and care have evolved the concept of “Care in Place,” a topic increasingly discussed in health care.¹⁹ Care in Place refers to the aim of reducing health disparities by providing hospital-level medical services and care to a patient within their own home. This attempt to decentralize medical care may involve a variety of interventions including wellness programs, telemedicine visits, home monitoring, mobile care service, and Smart home/Internet of Things.¹³

This shift toward Care in Place has implications for a variety of conditions and patient populations. Case studies or quasi experiments have especially focused on those patients for which transportation poses a great challenge, such as older patients²⁰ or those in palliative care,²¹ demonstrating the feasibility and efficacy of delivering hospital-quality care in the comfort of patients’ own homes. Care in Place has similarly been applied to the remote management of diabetic foot disease with a focus on digital health, Smart wearables, and telehealth.¹³ The “Wound Care Without Walls” (WCWW) initiative,²² proposed shortly after the beginning of the COVID-19 pandemic, has similarly aimed to conserve hospital resources. The WCWW “Pandemic Wound Triage System” involves triaging patients and providing as much care as possible in low-risk environments, only escalating the level of care when necessary to avoid subjecting patients to increased risk of exposure to COVID-19.

Patient Empowerment & Addressing Gaps in Chronic Care

Diabetes takes a disproportionate toll on communities of color,^4,5 as they experience more frequent and severe diabetes complications. These populations at greatest risk face barriers to managing their conditions, as well. Current technology for health management is often inaccessible for individuals with low literacy and low numeracy.^23,24 Provider bias and assumptions regarding potential usability and effectiveness of medical technology in minorities and other disadvantaged populations may pose yet another barrier to accessing high-quality care.^25,26

Thoughtful implementation of CV and ML within a mobile health system could aid decision-making processes during point of care interaction to improve overall effectiveness of clinical care. Armed with a more detailed and thorough understanding of their condition and current wound progress, patients may be better prepared to advocate for themselves during point of care interactions with their clinician(s). Especially in cases where a patient does not have the means to see the same clinician at regularly scheduled intervals, an increased understanding of their health allows patients to advocate for their own care.

Advantages of a mobile health system intervention include that it is not limited to clinician or clinic visits and that it also enables active engagement of informal caregivers and family members. While the Full Foot Exam is traditionally performed in a clinic annually (for individuals with diabetes but no history of DFUs) or in one- to four-week intervals (for individuals with DFUs), CV and ML approaches open the door for more frequent and standardized foot evaluation. Patients and their caregivers may monitor ulcers frequently and contact a clinician in the case of recurrence or change in severity. Furthermore, timeline images may be accessible via a clinician dashboard for remote monitoring of a patient’s DFU, allowing the clinical team to intervene in between routine visits if a change in care regimen is required.²⁷

By automating the Full Foot Exam and providing explanations and visualizations of foot health, CV and ML may act as vital computational tools to break down barriers to care and bring greater knowledge to the patients who need it most.¹³

CV for Visual Assessment of DFU

Wound surface area and depth estimates, part of the dermatologic component of the Full Foot Exam, are used to measure the DFU wound-healing progress. A ruler-based method is commonly employed by clinicians, but it can be time-consuming, subjective, and therefore inaccurate.^28,29 Given images of the diabetic foot, CV algorithms may automatically analyze images of the feet to segment feet and wounds (ulcers), and extract relevant characteristics, for example, measurements, shape, and location. This may also allow for precise documentation of wound size over time, aiding clinicians in the interpretation of a wound’s response to treatment.³⁰

Previous work¹⁵ has developed deep learning–based automated wound area assessment methods that integrate fully automatic wound segmentation, balancing computational efficiency and accuracy.³¹ Convolutional neural networks (CNNs) are a type of neural network which are designed for processing images and require a large, properly annotated data set. Such models are easily available off-the-shelf in popular programming languages such as Python (via the PyTorch framework) or may be downloaded from public repositories (such as HarDNet-DFUS³²). Given access to a large enough data set of annotated foot images, a CNN approach, such as Mask R-CNN, provides a simple, flexible, and general framework for object instance segmentation.³³ After the wound area is segmented, cardinal signs of inflammation (heat, redness, swelling, pain) and the classification of potential wound discharge could be leveraged for wound severity assessment. Wound attributes including images may be taken as input to predict clinical levels of infection severity.³⁴ Previously, CNN-based deep learning frameworks achieved F1 scores of over 80% when assessing wounds into five conditions using the eight wound attributes of the Photographic Wound Assessment Tool (PWAT): size, depth, necrotic tissue type, necrotic tissue amount, granulation tissue type, granulation tissue amount, edges, and periulcer skin viability.¹⁶ Wound depth and granulation tissues were also employed to grade wounds into five severities.³⁵

Table 1 provides a brief overview of common wound segmentation or classification model outputs and the associated CV or ML techniques used. Table 2 provides more detail on recent foot wound image studies in the literature,^15,16,34-42 including the participant cohort sizes and race/ethnicity, data set size (i.e., number of foot images), and CV or ML techniques used for wound segmentation or classification. Most DFU data sets that are currently publicly available were created to address binary classification of specific wound attributes, DFU detection, or DFU segmentation.²⁷ As noted in Table 2, very few of these studies include explicitly annotated infection or ischemia ground truth. Typically if these important indicators are included, it is only via a visual proxy such as yellow tissue^41,42 or they may be used during annotation as a medical indicator of the appropriate treatment decision.³⁴ The computational model of Shenoy et al⁴⁰ identifies the presence or absence of nine wound characteristics, including infection; however, they do not define their diagnostic criterion for infection. Given this gap, future work should develop models that use other nonvisual clinically relevant characteristics.

Table 1.

Automated Wound Segmentation and Classification at a Glance.

Common model outputs	Common CV/ML techniques
Wound segmentation
Segmentation mask Bounding box	YOLO, CNN, SVM YOLO, CNN, SVM
Wound classification
PWAT sub-scores Wound measurements or other characteristics	CNNCNN
Suggested treatment decision Wound healing indicator	SVM, Clustering SVM, Clustering
See Table 2 for examples.

Abbreviations: CNN, convolutional neural network; CV, computer vision; ML, machine learning; PWAT, Photographic Wound Assessment Tool; SVM, support vector machine; YOLO, You Only Look Once.

Table 2.

Overview of Studies Related to Automated Wound Segmentation and Classification.

Study	Data set				Model
	Participants		No. of images collected	Camera(s) & conditions	Model output	Ground truth annotations			CV/ML techniques
	No.	Race/ethnicity	No. of images collected	Camera(s) & conditions	Model output	Description	Infection	Ischemia	CV/ML techniques
Wound segmentation
Cassidy et al³⁶	Not stated	98% white 2% non-white	4200 images (DFUC 2020 data set)	Kodak DX4530, Nikon D3300, Nikon COOLPIX P100 Images captured in-clinic by medical photographers	Region detection: DFU bounding box	Wound bounding box annotated by three DFU-specializing healthcare professionals	No	No	YOLOv5; Faster R-CNN; EfficientDet object detection
Chairat et al³⁷	20	Not stated; all example photos are of white or pale skin	31 wound images	Smartphone Uncontrolled; collected in clinic; printed calibration chart (Macbeth chart, 4 ArUco markers)	Segmentation mask: wound area and wound tissue	Three physicians outlined wound area, granulation tissue, necrotic tissue, and epithelialization area	No	No	CNN (EfficientNet, U-Net with MobileNetV2)
Wang et al¹⁵	889	Not stated; all example photos are of white or pale skin	1109 foot ulcer images	Canon “SX 620 HS” digital camera, iPad Pro Uncontrolled illumination; various backgrounds	Segmentation mask of foot	Manual segmentation mask annotation, verified by wound care specialists	No	No	YOLOv3; CNN (MobileNetsV2)
Wang et al³⁸	15	Not stated; all example photos are of white or pale skin	100 foot ulcer images	Smartphone (Nexus 5) Controlled lighting and distance via capture box	Wound boundary segmentation mask	Three expert wound clinicians delineated wound boundaries	No	No	Support vector machine (SVM) classifier
Wound classification
Fong et al³⁹	82	Chinese (68%); Malay (12%); Indian (11%); Other (9%)	2334 venous leg ulcer images	Canon PowerShot G7X, Smartphone Uncontrolled, collected in clinic, reference ruler used	Wound measurements (length, width, and area)	Wound measurements recorded traditionally by clinician during visit	No	No	ML through “Tissue Analytics” application
Liu et al¹⁶	Not stated	Not stated; majority of example photos are of white or pale skin	1639 wound images (DFU, pressure, vascular, surgical)	Various (photos from medical archives; internet search; image capture box in clinic)	8 PWAT subscores, eg, size, depth, necrotic tissue size	Annotation of the eight PWAT subscores (methodology not stated)	No	No	CNN (DenseNet) with patch-based context-preserving attention mechanism
Nguyen et al³⁴	Not stated	Not stated	205 wound images (DFU, pressure, venous, arterial ulcers, surgical)	Various (corpus of patient data; internet search), controlled (pictures excluded if lighting too light or dark, or if only partial view of wound)	Appropriate treatment decision (#1 continue; #2 nonurgent change via wound specialist; #3 refer to specialist)	Annotation of the eight PWAT subscores (methodology not stated); Annotation of appropriate decision (#1, #2, or #3) by wound experts	Distant Proxy: Infection is a factor in treatment decision	Distant Proxy: Ischemia is a factor in treatment decision	Support Vector Machine (SVM) classifier, Gradient Boosted Machine (XGBoost)
Shenoy et al⁴⁰	Not stated	Not stated; all example photos are of white or pale skin	1335 wound images	Smartphone Uncontrolled; collected in clinic	Binary presence/absence of 9 characteristics, eg, wound, infection, staples, and drainage	Binary annotation of the 9 characteristics (Methodology not stated)	Binary presence/absence (diagnostic criterion not defined)	No	Multilabel CNN (VGG-16)
Wang et al^41,42	12	Not stated; all example photos are of white or pale skin	32 foot ulcer images	Smartphone Controlled lighting and distance via image capture box	“Wound Healing Score” based on area of wound (Red-Yellow-Black segmentation)	3 wound clinicians manually outlined wound area and scored healing status from 0 to 10	Proxy: yellow tissue indicates infection (not a clinically validated infection diagnostic criterion)	No	Mean-shift-based segmentation; K-means clustering
Zhao et al³⁵	Not stated	Not stated; all example photos are of white or pale skin	1639 diabetic wound images	Various, both controlled (collected with wound image box) and uncontrolled (internet search, collected in clinic)	Classify wound depth and granulation tissue on a 5-point scale	Wound depth and granulation tissue grade (based on PWAT scores) annotated in collaboration with wound experts	No	No	Bilinear CNN

Abbreviations: CNN, convolutional neural network; CV, computer vision; DFU, diabetic foot ulcer; ML, machine learning; PWAT, Photographic Wound Assessment Tool; SVM, support vector machine.

Many of the data sets used in previous deep learning implementations were collected in-clinic or under controlled conditions, and most of the studies referenced in Table 1 used data sets of primarily or entirely pale skin. An ML model is no better than its training data; biases present in the training data set can and do propagate to the model’s output. In particular, health disparities are propagated by the current, often-biased state of databases.⁴³ The equity of the deployment of these methods is threatened by algorithmic bias as well as a lack of transparency due to the difficulty of human interpretability. Further data collection of dark skin wounds is imperative to ensure future technologies will truly be equitable and address long-standing gaps in chronic care. Future tools must be implemented and evaluated in diverse populations, with a focus on those most heavily impacted by diabetes and DFUs.

ML Applications for Diabetic Foot

We now provide an overview of existing research in ML for the musculoskeletal domain of the Full Foot Exam, as well as predictive algorithms for diabetic foot care. We discuss not only the integration of multivariate information from heterogeneous data sources into a unified assessment framework but also computational modeling to automatically recognize early indicators of negative disease progression. Such predictive algorithms may be the key to breaking down barriers in care and bringing greater knowledge to patients in need.

ML for Diabetic Foot Gait Assessment

Changes in several gait characteristics are commonly associated with DFUs, such as gait speed, cadence, variability, and the percentage of cycle spent in the single-support phase.⁴⁴ Perhaps the most notable change, also associated with the neurological component of the Full Foot Exam, is the development of peripheral neuropathy.^44,45 When these peripheral nerves are damaged, patients with neuropathy may experience impaired sensation in the foot. Losing “the gift of pain” frequently delays clinical care of developing wounds, causing these patients to present with more advanced symptoms. Diabetes-related foot neuropathies lead to musculoskeletal foot deformities, biomechanical imbalances, foot ulcers, and ultimately delayed gait changes.^46-48 Automated gait analysis may be used to track mechanical changes or conservative gait strategy due to pain, calluses, swelling, or peripheral neuropathy. Increased risk of injury may be identified even before any visible changes are present, enabling proactive intervention and potential changes to care regimens.

The technological key to automated gait monitoring is advanced signal processing methods and time-series analysis methods for change point detection. Previous systems used pressure sensor flooring,⁴⁹ or inertial measurement units (IMUs) attached to the lower back, lumber levels, tibias, or feet.^45,50-52 Some previous work has relied solely on IMU data (triaxial accelerometer and gyroscope data) from a common smartphone for gait assessment. This work frequently requires standardized positioning of the smartphone’s internal IMU by holstering the phone in a belt pouch⁵³ or a trouser pocket.^54,55

Many existing gait databases, such as the OU-MVLP database,⁵⁶ are not specific to those with gait disorders. In addition, there is high comorbidity between DFUs, obesity, and aging. Machine learning gait assessment must be able to discriminate characteristics that are indicative of progressing DFUs from those of progressing age or complications with weight. Existing work has investigated the impacts of obesity and aging on gait stability and variability.^57-59

Predictive ML Algorithms for Diabetic Foot Care

Given a heterogeneous data set from patients with DFU including data streams from the components of the Full Foot Exam, an ML model may be robustly trained and evaluated through leave-one-out cross-validation protocols. Previous work showed high predictive ability for non-amputation, minor amputation, or major amputation in hospitalized patients with DFUs by leveraging five-fold cross-validation based on a variety of patient data including the presence of ischemia and/or infection.¹⁷ Similarly, an artificial neural network provided accuracy of 82% when predicting the healing outcomes of patients with DFU based on six clinical features and supported the creation of a simple wound healing prediction software tool.¹⁸ The heterogeneous data may also be used to train generalizable assessment methods based on multivariate sequence modeling approaches, which attempt to make inferences about temporal sequences of data points. For example, long short-term memory (LSTM)⁶⁰ is a neural network architecture which includes a memory mechanism, allowing it to model long-term dependencies. These approaches may be combined with explicit representation models, such as variational auto-encoders⁶¹ to extract a useful and representative set of features for each datapoint. In addition, any of these assessment methods may be pretrained on aggregate patient data and fine-tuned (adapted) toward the idiosyncrasies of individual patients.

In addition to the overall assessment of disease progression, computational modeling may provide an opportunity to automatically recognize early indicators of negative disease progression via sequence forecasting (see Shi and Yeung⁶² for a survey), based on previous work on attention-modeling in deep learning-based time-series analysis.⁶³ Long short-term memory sequential models of sensor data streams and attention weight distributions may be trained not only for each sensing modality but also for each timestep in a sequence of measurements.⁶⁴ Through the automated analysis of measurements (both modality and timestep) that primarily contribute to the overall assessment, models may be derived to predict the outcome of the DFU assessment sooner than four weeks after the last medical visit.

These approaches integrate multivariate information from heterogeneous data sources into a unified assessment framework. New data aggregation models will be required that cope with multimodality, sparsity, and potential ambiguity of the input sequences. Embedding methods⁶⁵ learn combined feature spaces directly from input data sources and cluster semantically coherent portions of the input data. Feature embeddings represent the state-of-the-art in domains such as natural language processing. Preliminary results have been reported for more signal-oriented domains, eg, acoustic signal embeddings.⁶⁶ Multivariate embeddings address sparsity issues and allow for the extraction of essential trajectories of DFU disease progression that may then be used for forecasting and early prediction. Semantic embedding facilitates effective data imputation. Attention modeling facilitates effectiveness analysis of the overall procedure, eg, by eliminating data points that do not contribute to the overall assessment result.

Limitations and Future Work

Given the complexity of foot ulceration and frequent recurrence, diabetic foot care may benefit greatly from the development of further CV and ML approaches to detect wounds and monitor their progression. Future work might consider the role of predictive ML applications that incorporate additional patient medical data such as A1c or X-rays of the foot.

The use of CV is limited by multiple factors. Computer vision models require large-scale data annotation and ground truth verification by clinical staff, which can be labor-intensive with prohibitive cost. In addition, CV is primarily based on images and is therefore limited to what can be inferred visually. More sophisticated data capturing devices such as a depth camera or thermal camera may be able to assess other clinically relevant characteristics such as depth, or infection and ischemia (without relying on a proxy indicator such as tissue color), but they are expensive and would require novel ML techniques. Another promising venue is to use video to investigate CV techniques for the capillary refill test (CRT) within the vascular domain of the Full Foot Exam. In practice, the recorded output of the CRT is binary, less than three seconds or greater than three seconds, and is not a reliable predictor of ischemia, which is an important indicator for both DFU occurrence and recurrence.⁶⁷ The incorporation of CV for automated analysis of a CRT video may allow for higher precision results even if invisible to the naked eye.

Individuals are best understood when viewed as the center of an ecological system,⁶⁸ taking into account the intersectional and multifaceted challenges of DFU management for all stakeholders: patient, caregiver(s), family, community, society, and environment. This ecological approach steers us away from a more traditional view of chronic care centered only around routine and relatively infrequent visits to clinicians and instead highlights the need for collaborative telecare, improvements to patient-clinician communication and patient-caregiver interaction, patient-gathered data, education, and engagement. Current technology for health management is often not user-friendly for individuals with low literacy and low numeracy.^23,24 Design of DFU technologies must consider personal, structural, and environmental factors^4,68 due to the challenges faced during real-world deployments, as explored in previous mobile health technology studies for diabetic foot care.^69-71 In many cases, individuals with diabetes learn about proper footcare only after developing complications, instead of earlier in the course of their disease.⁷² Thus, future work should take an ecological approach regarding three key elements that underpin efforts to prevent foot ulcers and amputations: identifying the at-risk foot; regularly promoting inspection and examination of the at-risk foot, and supporting the engagement and education of the patient, caregiver(s), family, community, and care professionals in language and methods appropriate for their needs and abilities.

Computer vision and ML offer a unique opportunity to bolster patient empowerment by providing more concrete measures and clear visualizations of wound health, as well as interpretable analyses of a patient’s health data over time. Remote monitoring and automated classification are expected to revolutionize remote wound care by allowing patients to self-monitor their wound pathology, assist in the remote triaging of patients by clinicians, and allow for more immediate interventions when necessary.⁷³ However, there does not currently exist a computational model that analyzes the heterogeneous biomedical data involved in the Full Foot Exam, nor does there exist an associated interface that connects patients and clinicians and assists in interpreting this data. We contribute this overview of current CV and ML literature and its applicability to diabetic foot care to provide a basis for future work in these areas.

Footnotes

Abbreviations

CNN, convolutional neural network; CRT, capillary refill test; CV, computer vision; DFU, diabetic foot ulcer; IMU, inertial measurement unit; LSTM, long short-term memory; ML, machine learning; PWAT, Photographic Wound Assessment Tool; SVM, support vector machine; WCWW, “Wound Care Without Walls”; YOLO, “You Only Look Once” object detection.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the NIH Exploratory/Developmental Research Grant Award R21 (grant no. 1R21MD017943-01 and 5K23DK124647-03) and the GT-Emory AI. Humanity seed grant.

ORCID iDs

Cynthia Baseman

Maya Fayfman

Marcos Schechter

Gabriel Santamarina

Thomas Ploetz

Rosa I. Arriaga

References

About prediabetes and type 2 diabetes. https://www.cdc.gov/diabetes/prevention/about-prediabetes.html

Thorpe

Ogden

Galactionova

. Chronic conditions account for rise in Medicare spending from 1987 to 2006. Health Aff. 2010;29(4):718-724.

Armstrong

Boulton

AJM

Bus

. Diabetic foot ulcers and their recurrence. N Engl J Med. 2017;376:2367-2375.

Minority health at HHS. https://minorityhealth.hhs.gov/omh/browse.aspx?lvl=4&lvlid=18

Murphy

Kochanek

Arias

Tejada-Vera

. Deaths: final data for 2018. Natl Vital Stat Rep. 2021;69(13):1-83.

Harding

Andes

Rolka

, et al. National and state-level trends in nontraumatic lower-extremity amputation among U.S. Medicare beneficiaries with diabetes, 2000–2017. Diabetes Care. 2020;43(10):2453-2459.

McDermott

Fang

Boulton

, et al. Etiology, epidemiology, and disparities in the burden of diabetic foot ulcers. Diabetes Care. 2023;46(1):209–221.

Faglia

Clerici

Clerissi

, et al. Long-term prognosis of diabetic patients with critical limb ischemia: a population-based cohort study. Diabetes Care. 2009;32(5):822-827.

Boulton

AJM

Armstrong

Albert

, et al. Comprehensive foot examination and risk assessment. A report of the Task Force of the Foot Care Interest Group of the American Diabetes Association, with endorsement by the American Association of Clinical Endocrinologists. Phys Ther. 2008;88:1436-1443.

10.

Miller

Carter

Shih

, et al. How to do a 3-minute diabetic foot exam. J Fam Pract. 2014;63:646-656.

11.

Lipsky

Senneville

Abbas

, et al. Guidelines on the diagnosis and treatment of foot infection in persons with diabetes (IWGDF 2019 update). Diabetes Metab Res Rev. 2020;36(suppl 1):e3280.

12.

Russell

Norvig

. Artificial Intelligence: A Modern Approach. 4th ed. Berkeley: University of California, Berkeley; 2021.

13.

Najafi

Mishra

. Harnessing digital health technologies to remotely manage diabetic foot syndrome: a narrative review. Medicina. 2021;57:377. doi:10.3390/medicina57040377.

14.

Ossai

Wickramasinghe

. Automatic user sentiments extraction from diabetes mobile apps—an evaluation of reviews with machine learning. Inform Health Soc Care. 2023;48:211-230. doi:10.1080/17538157.2022.2097083.

15.

Wang

Anisuzzaman

Williamson

, et al. Fully automatic wound segmentation with deep convolutional neural networks. Sci Rep. 2020;10:21897.

16.

Liu

Agu

Pedersen

Lindsay

Tulu

Strong

. Comprehensive assessment of fine-grained wound images using a patch-based CNN with context-preserving attention. IEEE Open J Eng Med Biol. 2021;2:224-234.

17.

Xie

Deng

, et al. An explainable machine learning model for predicting in-hospital amputation rate of patients with diabetic foot ulcer. Int Wound J. 2022;19(4):910-918.

18.

Poradzka

Czupryniak

. The use of the artificial neural network for three-month prognosis in diabetic foot syndrome. J Diabetes Complications. 2023;37(2):108392.

19.

Ivanova

Wallenburg

Bal

. Care in place: a case study of assembling a carescape. Sociol Health Illn. 2016;38(8):1336-1349.

20.

Leff

Burton

Mader

, et al. Hospital at home: feasibility and outcomes of a program to provide hospital-level care at home for acutely ill older patients. Ann Intern Med. 2005;143:798-808.

21.

Laranjeira

Dixe

Martinho

Rijo

Querido

. Building bridges for “palliative care-in-place”: development of a mHealth intervention for informal home care. Front Psychol. 2022;13:862347.

22.

Rogers

. How COVID-19 is changing wound care and why you should be preparing now for the next wave. June 2020. https://www.hmpgloballearningnetwork.com/site/twc/articles/how-covid-19-changing-wound-care-and-why-you-should-be-preparing-now-next-wave#:~:text=A%20group%20of%20wound%20care%20leaders%20and%20innovators,the%20utilization%20and%20site%20of%20care%20%28Figure%201%29

23.

Barnes

Caine

Connelly

Siek

. Understanding the needs of low SES patients with type 2 diabetes. In: 2013 7th International Conference on Pervasive Computing Technologies for Healthcare and Workshops; 2013:302-306. https://ieeexplore.ieee.org/document/6563950

24.

Diethei

Singh

Wienert

, et al. Learning from mHealth success stories: effective interventions for marginalised populations. In: 22nd International Conference on Human-Computer Interaction With Mobile Devices and Services. New York, NY: Association for Computing Machinery; 2021:1-4. https://dl.acm.org/doi/10.1145/3406324.3424653

25.

Agarwal

Schechter

Gonzalez

Long

. Racial-ethnic disparities in diabetes technology use among young adults with type 1 diabetes. Diabetes Technol Ther. 2021;23:306-313.

26.

Agarwal

Crespo-Ramos

Long

Miller

. “I didn’t really have a choice”: qualitative analysis of racial-ethnic disparities in diabetes technology use among young adults with type 1 diabetes. Diabetes Technol Ther. 2021;23(9):616-622. doi:10.1089/dia.2021.0075.

27.

Yap

Kendrick

Reeves

Goyal

Pappachan

Cassidy

. Development of diabetic foot ulcer datasets: an overview. In: Diabetic Foot Ulcers Grand Challenge. Springer; 2022:1-18. https://arxiv.org/abs/2201.00163

28.

Foltynski

Ladyzynski

Ciechanowska

Migalska-Musial

Judzewicz

Sabalinska

. Wound area measurement with digital planimetry: improved accuracy and precision with calibration based on 2 rulers. PLoS One. 2015;10(8):e0134622.

29.

Shah

Wollak

Shah

. Wound measurement techniques: comparing the use of ruler method, 2D imaging and 3D scanner. J Am Coll Clin Wound Spec. 2013;5(3):52-57.

30.

Goyal

Reeves

Rajbhandari

Yap

. Computerised methods for monitoring diabetic foot ulcers on plantar foot: a feasibility study. In: Medical Image Understanding and Analysis. Springer; 2022:199-211. https://dl.acm.org/doi/10.1007/978-3-031-12053-4_15

31.

Sandler

Howard

Zhu

Zhmoginov

Chen

L-C

. MobileNetV2: inverted residuals and linear bottlenecks. In: 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE; 2018:4510-4520. https://ieeexplore.ieee.org/document/8578572

32.

Liao

Yang

Lai

Shen

Lin

. HarDNet-DFUS: an enhanced harmonically-connected network for diabetic foot ulcer image segmentation and colonoscopy polyp segmentation. arXiv. 2022. https://arxiv.org/abs/2209.07313

33.

Gkioxari

Dollár

Girshick

. Mask R-CNN. In: Proceedings of the IEEE International Conference on Computer Vision; 2017:2961-2969. https://ieeexplore.ieee.org/document/8237584

34.

Nguyen

Agu

Tulu

, et al. Machine learning models for synthesizing actionable care decisions on lower extremity wounds. Smart Health. 2020;18:100139. doi:10.1016/j.smhl.2020.100139.

35.

Zhao

Liu

Agu

, et al. Fine-grained diabetic wound depth and granulation tissue amount assessment using bilinear convolutional neural network. IEEE Access. 2019;7:179151-179162.

36.

Cassidy

Reeves

Pappachan

, et al. The DFUC 2020 dataset: analysis towards diabetic foot ulcer detection. touchREV Endocrinol. 2021;17(1):5-11.

37.

Chairat

Chaichulee

Dissaneewate

Wangkulangkul

Kongpanichakul

. AI-assisted assessment of wound tissue with automatic color and measurement calibration on images taken with a Smartphone. Healthcare. 2023;11:273. doi:10.3390/healthcare11020273.

38.

Wang

Pedersen

Agu

Strong

Tulu

. Area determination of diabetic foot ulcer images using a cascaded two-stage SVM-based classification. IEEE Trans Biomed Eng. 2017;64(9):2098-2109.

39.

Fong

Lai

Chan

, et al. Clinical validation of a smartphone application for automated wound measurement in patients with venous leg ulcers. Int Wound J. 2023;20(3):751-760.

40.

Shenoy

Foster

Aalami

Majeed

Aalami

. Deepwound: automated postoperative wound assessment and surgical site surveillance through convolutional neural networks. In: 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM); 2018:1017-1021. https://arxiv.org/abs/1807.04355#:~:text=11%20Jul%202018%5D-,Deepwound%3A%20Automated%20Postoperative%20Wound%20Assessment%20and%20Surgical,Surveillance%20through%20Convolutional%20Neural%20Networks&text=Postoperative%20wound%20complications%20are%20a,wound%20complications%20is%20strongly%20desired

41.

Wang

Pedersen

Strong

, et al. An automatic assessment system of diabetic foot ulcers based on wound area determination, color segmentation, and healing score evaluation. J Diabetes Sci Technol. 2015;10:421-428.

42.

Wang

Pedersen

Strong

Tulu

Agu

Ignotz

. Smartphone-based wound assessment system for patients with diabetes. IEEE Trans Biomed Eng. 2015;62(2):477-488. doi:10.1109/tbme.2014.2358632.

43.

Norori

Aellen

Faraci

Tzovara

. Addressing bias in big data and AI for health care: a call for open science. Patterns. 2021;2(10):100347.

44.

Courtemanche

Teasdale

Boucher

Fleury

Lajoie

Bard

. Gait problems in diabetic neuropathic patients. Arch Phys Med Rehabil. 1996;77(9):849-855.

45.

Esser

Collett

Maynard

, et al. Single sensor gait analysis to detect diabetic peripheral neuropathy: a proof of principle study. Diabetes Metab J. 2018;42(1):82-86.

46.

Arvanitakis

Wilson

Schneider

Bienias

Evans

Bennett

. Diabetes mellitus and progression of rigidity and gait disturbance in older persons. Neurology. 2004;63:996-1001.

47.

Pakarinen

Laine

Honkonen

Peltonen

Oksala

Lahtela

. Charcot arthropathy of the diabetic foot. Current concepts and review of 36 cases. Scand J Surg. 2002;91(2):195-201.

48.

Rogers

Bevilacqua

. The diagnosis of Charcot foot. Clin Podiatr Med Surg. 2008;25:43-51, vi.

49.

Wuehr

Schniepp

Schlick

, et al. Sensory loss and walking speed related factors for gait alterations in patients with peripheral neuropathy. Gait Posture. 2014;39(3):852-858.

50.

Awotidebe

Ativie

Oke

, et al. Relationships among exercise capacity, dynamic balance and gait characteristics of Nigerian patients with type-2 diabetes: an indication for fall prevention. J Exerc Rehabil. 2016;12(6):581-588.

51.

Watanabe

Noguchi

Sanada

Mori

. Development of a plantar load estimation algorithm for evaluation of forefoot load of diabetic patients during daily walks using a foot motion sensor. J Diabetes Res. 2017;2017:5350616.

52.

Esser

Dawes

Collett

Howells

. Insights into gait disorders: walking variability using phase plot analysis, Parkinson’s disease. Gait Posture. 2013;38(4):648-652. doi:10.1016/j.gaitpost.2013.02.016.

53.

Muaaz

Mayrhofer

. An analysis of different approaches to gait recognition using cell phone based accelerometers. In: Proceedings of International Conference on Advances in Mobile Computing & Multimedia. New York, NY: Association for Computing Machinery; 2013:293-300. https://dl.acm.org/doi/10.1145/2536853.2536895

54.

Kao

H-L (Cindy)

B-J

Lin

Chu

H-H

. Phone-based gait analysis to detect alcohol usage. In: Proceedings of the 2012 ACM Conference on Ubiquitous Computing. New York, NY: Association for Computing Machinery; 2012:661-662. https://oa.mg/work/10.1145/2370216.2370354

55.

Thang

Viet

Dinh Thuc

Choi

. Gait identification using accelerometer on mobile phone. In: 2012 International Conference on Control, Automation and Information Sciences (ICCAIS); 2012:344-348. https://ieeexplore.ieee.org/document/6466615

56.

Takemura

Makihara

Muramatsu

Echigo

Yagi

. Multi-view large population gait dataset and its performance evaluation for cross-view gait recognition. IPSJ Trans Comput Vis Appl. 2018;10:4.

57.

Gonzalez

Gates

Rosenblatt

. The impact of obesity on gait stability in older adults. J Biomech. 2020;100:109585.

58.

da Silva-Hamu

TCD

Formiga

CKMR

Gervásio

Ribeiro

Christofoletti

de França Barros

. The impact of obesity in the kinematic parameters of gait in young women. Int J Gen Med. 2013;6:507-513.

59.

Callisaya

Blizzard

Schmidt

McGinley

Srikanth

. Ageing and gait variability—a population-based study of older people. Age Ageing. 2010;39(2):191-197.

60.

Hochreiter

Schmidhuber

. Long short-term memory. Neural Comput. 1997;9:1735-1780.

61.

Gan

Henao

, et al. Variational autoencoder for deep learning of images, labels and captions. In: Advances in Neural Information Processing Systems 29 (NIPS 2016); 2016. https://proceedings.neurips.cc/paper/2016/hash/eb86d510361fc23b59f18c1bc9802cc6-Abstract.html

62.

Shi

Yeung

D-Y

. Machine learning for spatiotemporal sequence forecasting: a survey. arXiv. 2018. http://arxiv.org/abs/1808.06865

63.

Murahari

Plötz

. On attention models for human activity recognition. In: Proceedings of the 2018 ACM International Symposium on Wearable Computers. New York, NY: Association for Computing Machinery; 2018:100-103. https://dl.acm.org/doi/10.1145/3267242.3267287

64.

Zeng

Gao

, et al. Understanding and improving recurrent networks for human activity recognition by continuous attention. In: Proceedings of the 2018 ACM International Symposium on Wearable Computers. New York, NY: Association for Computing Machinery; 2018:56-63. https://dl.acm.org/doi/10.1145/3267242.3267286

65.

Kusner

Sun

Kolkin

Weinberger

. From word embeddings to document distances. In: Bach

Blei

, eds. Proceedings of the 32nd International Conference on Machine Learning. Lille, France: PMLR; 2015:957-966.

66.

Chung

Y-A

Glass

. Speech2Vec: a sequence-to-sequence framework for learning word embeddings from speech. arXiv. 2018. http://arxiv.org/abs/1803.08976

67.

Meloni

Izzo

Giurato

Lázaro-Martínez

Uccioli

. Prevalence, clinical aspects and outcomes in a large cohort of persons with diabetic foot disease: comparison between neuropathic and ischemic ulcers. J Clin Med Res. 2020;9:1780. doi:10.3390/jcm9061780.

68.

Bronfenbrenner

. Ecological systems theory. In: Vasta

, ed. Six Theories of Child Development: Revised Formulations and Current Issues. London, England: Jessica Kingsley Publishers; 1992:187-249.

69.

Anthony

Femino

Miller

, et al. Diabetic foot surveillance using mobile phones and automated software messaging, a randomized observational trial. Iowa Orthop J. 2020;40(1):35-42.

70.

Swerdlow

Shin

D’Huyvetter

Mack

Armstrong

. Initial clinical experience with a simple, home system for early detection and monitoring of diabetic foot ulcers: the foot selfie. J Diabetes Sci Technol. 2023;17(1):79-88.

71.

Kong

Ramirez-GarciaLuna

Fraser

RDJ

Wang

. A 57-year-old man with type 1 diabetes mellitus and a chronic foot ulcer successfully managed with a remote patient-facing wound care Smartphone application. Am J Case Rep. 2021;22:e933879.

72.

Fayfman

Schechter

Amobi

, et al. Barriers to diabetic foot care in a disadvantaged population: a qualitative assessment. J Diabetes Complications. 2020;34(12):107688.

73.

Pappachan

Cassidy

Fernandez

Chandrabalan

Yap

. The role of artificial intelligence technology in the care of diabetic foot ulcers: the past, the present, and the future. World J Diabetes. 2022;13:1131-1139.